应用模糊技术的递归小波神经网络太阳日总辐射预测

Forecast of Daily Total Solar Irradiance Using Recurrent Wavelet BP Neural Networks Assisted by Fuzzy Technology

考虑到人工神经网络在非线性函数逼近方面的特性和小波函数良好的时频域多分辨分析能力,建立了结合两者优点的递归小波BP网络(RWBPNN)模型,用以对地面太阳日总辐射进行准确预测.该模型将气象台的天气阴晴预报进行模糊化处理后输入神经网络,增加有用信息以改善模型的预测精度.同时还提出了批量平均权值法来训练网络,有效地改善了初始参数的选择问题.实例以及模型间的比较说明了本模型应用于太阳辐射预测具有更高精度和实际可行性.

In consideration of the good performance of artificial neural networks in approximating nonlinear functions and the prominent ability of wavelet functions in time-frequency domain multi-resolution analyses, a recurrent wavelet BP neural network （RWBPNN） is established in combination of both advantages so as to forecast exactly the daily total solar irradiance. In order to further improve the forecast precision, the cloudiness from weather forecast is fuzzilized and then to be the inputs of the RWBPNN. A batch-average-weights method is used in network training for more effective selection of initial parameters. As an example a daily irradiance forecast for a month is completed using the sample data in Macao, and comparisons between irradiation models show that the RWBPNN model has higher precisions and is more feasible.